Co-Clustering by Directly Solving Bipartite Spectral Graph Partitioning.

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Bipartite spectral graph partitioning (BSGP) is a co-clustering method for simultaneous document and word clustering.
• Traditional BSGP relies on graph construction and singular value decomposition, which can be suboptimal.
• Existing methods face challenges with unstructured, fixed graphs and potential deviations in the two-stage process.

Purpose of the Study:

• To propose a novel bipartite graph partitioning method that directly learns a graph structure.
• To address the limitations of traditional BSGP methods, including graph quality and process deviation.
• To achieve direct clustering results by learning a bipartite graph with a specified number of connected components.

Main Methods:

• A new bipartite graph partitioning method is proposed.
• The method learns a bipartite graph with exactly 'c' connected components, where 'c' is the number of clusters.
• The approach aims to provide direct clustering results without a separate post-processing step.

Main Results:

• The proposed model's solution is theoretically and experimentally proven to be the discrete solution of the primal BSGP problem in specific cases.
• Experimental results on synthetic and real-world datasets demonstrate the superiority of the proposed method.
• The novel method achieves direct clustering and enhances the quality of co-clustering.

Conclusions:

• The developed bipartite graph partitioning method offers a significant improvement over traditional BSGP.
• The direct learning of graph structure leads to more accurate and efficient co-clustering.
• The method shows strong performance and potential for document clustering applications.